Papers

Long-horizon agents need to remember what happened over weeks and months, yet today's external memory systems either store flat recent history or perform destructive one-shot compaction that discards the detail needed for evidence and reasoning. We present Chronomem ( 忆 时 塔 , "Tower of Remembering Time"), a time-layered external memory service exposed through the Model Context Protocol (MCP). Inspired by two everyday mechanisms that roll detail up layer by layer—financial consolidation and the daily-report chain (task → daily → weekly → monthly)—Chronomem organizes episodic records into a hierarchy of time windows—from ten minutes up to a week—where each level is a materialized snapshot that references finer levels and can be drilled down to original records. Its central design principle is the separation of two kinds of recall with different costs: exact recall, which navigates reference chains to raw records and costs no LLM tokens, and semantic recall, which reads LLMgenerated summaries and is the form that truly compresses coarse windows. Summary records are themselves append-only temporal entries, so summaries can be recursively summarized, realizing progressive, non-destructive compression. Memory types are user-defined with per-type granularity chains, so high-frequency channels (chat, tool calls) keep minute-level granularity while sparse channels (work logs) converge to day/week. We position the design, discuss open problems (semantic drift, token-budgeted trigger strategies, privacy), and outline an evaluation plan; empirical results are forthcoming.

This document presents a rebuttal to the paper “ChatGPT is bullshit” (CiB), which claims that Large Language Models (LLMs) like ChatGPT are inherently unconcerned with truth, merely imitating human-like text without the capacity for genuine intention. Our counterargument emphasizes four main points: the critical role of Reinforcement Learning from Human Feedback (RLHF), the conflation of specific domain failures with a lack of overall truth-orientation, the false dichotomy of “truth-valuing humans vs. bullshitting LLMs”, and the risk that CiB itself meets its own criteria for “bullshit”. We conclude with reflections on how responsible, transparent, and up-to-date research is essential for understanding the evolving capacities of AI systems. This text was coauthored by a human and ChatGPT (o1 version), following a transparent collaborative process documented in the appendices. A noteworthy point is that o1, unprompted, identifies the critical error in CiB’s argument after only a few seconds of thought.

"Vibe coding" describes a style of software development in which a human works interactively with an AI coding assistant, specifying goals and evaluating results while delegating much of the detailed implementation, testing, debugging, and explanation to the AI. Similar methods can now be applied in the academic world. We suggest the term "vibe research" for sustained research collaboration of this kind. The analogy raises an immediate puzzle. If vibe coding and vibe research involve closely related forms of human-AI collaboration, why has vibe coding already become a generally accepted software development practice, while vibe research remains difficult to accommodate within mainstream academia? We consider three possible explanations. First, even if vibe research is technically possible, AI contributions may be too limited or trivial for the question to matter in practice. Second, research collaboration may require forms of responsibility or accountability which AI systems cannot assume. Third, current academic institutions may simply not provide a place for this form of collaboration within their established governance structures. We argue that each explanation is less straightforward than it initially appears. With regard to the substantive importance of AI-generated results, we cite multiple recent examples from mathematics in which frontier AI systems have made contributions regarded by the mathematical research community as highly non-trivial. With regard to responsibility, it is indeed true that an AI cannot assume institutional accountability in the same way as a human. We argue, however, that for many of the questions at issue the operationally useful requirement is reliable intellectual provenance: the ability to establish where ideas and decisions came from, inspect the evidence on which they were based, and obtain an explanation of them. The infrastructure used to preserve project history, trace decisions, inspect artefacts, and explain technical choices in vibe-coding projects carries over naturally to vibe-research projects and can deliver this kind of provenance. We present four examples from our own work showing that this is not merely a theoretical possibility but can readily be realised in practice. In the fourth, the present paper itself, we go beyond preserving final research artefacts and supporting resources: the complete human-AI session trace through which the paper was developed is retained as part of the provenance record, together with lightly processed representations and metadata supporting navigation and search. With regard to institutional acceptance, we argue that an important sociological difference between the software and academic worlds is that academic publishing is governed by formal authorship and ethics policies which change through relatively slow institutional processes, while software-development methodology can often change directly through team and organisational practice. We examine the ACL Policy on Publication Ethics as a particularly relevant example. Our general aims in the paper are not to argue that vibe research is risk-free, nor to propose a general rule for AI authorship. They are rather to identify an apparent asymmetry between software engineering and academic research, examine some of its conceptual, technical, and institutional foundations , and ask whether current academic conventions accurately describe emerging forms of human-AI research collaboration.

以人工智能与人形机器人为核心的新一轮技术变革，不同于历史上历次工业革命，它能够将人类成熟生产经验封装入自动化系统，实现 24 小时持续生产，大幅压缩制造业产品边际成本，引发行业竞争重构与结构性失业风险。马斯克进一步提出配套能源与算力方案：依托太空轨道太阳能近乎不间断捕获光能，将高能耗 AI 算力、生产经验封装、模型推理任务部署在太空轨道完成，仅把最终结果传回地球，构建 “太空供能 + 太空算力 + 地面自动化生产” 的商业闭环，从根源缓解地面 AI 算力电力瓶颈，打破化石能源约束，进一步放大物质生产的潜力。马斯克提出 AI 与机器人将带来物质极度富足的未来，该论断描述了生产力提升带来的可能性边界，但并不等于必然现实。本文论证：技术本身具备中性属性，它可以放大生产力、拓宽人类闲暇，也可以放大原有社会分配失衡；技术只会带来富足的可能性，真正实现普惠富足，取决于社会制度、分配机制，以及社会个体的自主选择。当生产红利集中于少数资本与技术所有者，而社会保障、再分配体系滞后时，技术进步会加剧阶层分化，激化社会矛盾；若配套合理的分配改革、终身学习体系与普惠保障，则技术释放的生产力可以转化为全体社会成员的福利提升。

本文以烹饪活动作为具象隐喻，探讨人类认知生成、AI 技能封装以及人形具身智能的演进逻辑，辨析当前大语言模型的固有能力边界。人类认知根植于实操试错，在持续反馈循环中沉淀经验；数字大模型可将标准化经验封装为 Skill 技能包，实现流程复用，降低重复劳动成本，但尚不具备自主范式跃迁能力。产业技术演进存在时序差异：数字智能体优先接管数字空间任务；后续人形机器人与具身智能将面向物理世界落地，把人类实操经验转化为具身执行技能。物理环境存在大量动态不确定性，使得物理场景下的技能封装难度显著高于数字场景。依托「爱因斯坦测试」的复古大模型实测研究：仅使用相对论诞生之前的全部知识训练大模型，模型无法独立完整推导出相对论，证明当前大模型仅能在已有知识体系内归纳拟合，无法自主完成范式级的底层理论突破。研究表明，实操催生认知，标准化经验能够封装复用；AI 与具身机器人擅长复刻、执行人类已沉淀完成的经验；范式级的认知突破仍是人类独有的能力，人机协同将成为未来主流工作模式。

本文提出了一种面向中国独立研究者需求量身定制的新型预印本平台模型。与现有需要机构隶属关系或背书的平台不同，拟建系统结合了两个核心功能：（1）零门槛预印本发布；（2）社区驱动的讨论与协作。该平台集成了一个AI辅助质量验证系统，对提交的稿件进行多轮基于对话的验证。同时设计了一套分级用户成长机制，逐步将有潜力的独立研究者与知名学者和机构连接起来。本方案旨在弥合中国日益增长的研究产出与其对独立研究者有限的开放获取基础设施之间的差距

By early 2026, the discourse around agent harnesses — the configuration layer of skills, rules, prompts, and documentation shaping an LLM agent's behavior — has named two activities: harness engineering, the reactive practice of ensuring an agent never repeats a mistake, and harness optimization, autonomous search over harness code for benchmark score. Both run against fixed, checkable criteria. The activity on the other side — keeping the harness aligned with what its operator now wants, where the criterion moves — is performed daily and has, to the author's knowledge, no established name. The paper defines harness alignment — the continuous, human-gated activity of keeping an agent's harness aligned with the operator's evolving intent — and its failure mode, harness drift. The three defining properties — continuous, human-gated, bidirectional — follow from a single root: intent, unlike correctness, cannot be automated the same way — it has no verifier outside the operator, and moves as the operator's judgment sharpens; verifying intent sharpens the judgment doing it, so the loop moves its own target. An automated check freezes intent into a specification, reducing its automatable part to correctness work. A four-domain search found no established term covering all three; an audit shows 2026 drift coinages severed from the classical lineage that harness drift bridges. A six-phase cycle (Research, Extract, Curate, Promote, Measure, Maintain) operationalizes it over a memory structure whose boundary separates freely-writable records from gated behavior-shaping artifacts. Failure has two layers: artifact-side harness drift, and a human-side twin — gate complacency, deskilling, delegation-feedback divergence — anchored in the automation-complacency and ironies-of-automation literatures — structural inference, not measurement. Two running instances, differing in substrate, model, and knowledge genre, are offered as portability evidence, not efficacy. All of it comes from months-old practice — provisional judgments offered for testing, not settled practice.

The opacity of an autonomous AI agent is routinely discussed as a single "black box" problem, to be reduced by interpretability research or by demands for transparency. This paper argues that the opacity has two architecturally distinct causes that the discourse conflates, and that the conflation is what makes the transparency debate unwinnable. One cause is technical: behavior internalized into model weights cannot be read, versioned, or reverted without retraining. The other is commercial: the human-built layer around the weights — system prompts, rules, tool definitions, the agent loop — is technically inspectable but kept proprietary because it is a competitive differentiator. The two causes admit different responses, and treating them as one problem routes every transparency argument into an irresolvable safety-versus-secrecy collision. The paper makes three contributions, drawn from the implementation history of a single agent and re-expressed as harness-neutral judgments. First, it grounds the opacity problem in an *accountability gap*: a text-based prohibition on a capability that physically exists is a signpost, not a control, and enforcement admits a *prohibition-strength hierarchy* — absence, then scaffolding-layer enforcement, then untrusted-content boundary — most published safety work occupying only the weakest layer. Second, it separates the *two-layer black box*: technical internalization (Layer 1, largely permanent) from commercial secrecy of scaffolding (Layer 2, contingent on market structure). Third, it resolves Layer 2 with a *minimum disclosure set*: operator visibility is decoupled from public disclosure, and the minimum that makes post-incident causal tracing possible — which scaffolding version was active, which inputs reached the agent, which approval gated the version — is disclosable without publishing the scaffolding text. The resolution is implementable with current technology; it waits on neither an interpretability breakthrough nor a restructuring of markets and law. It complements existing AI risk-management and management-system standards by naming the operator-visibility floor those standards presuppose. The open questions — where exactly the disclosure line falls, and how the technology–society timescale gap is to be managed — are the research agenda.

Autonomous AI agents in business deployments exhibit a recurring failure mode: when an incident occurs, responsibility cannot be redirected to a separable contributor. The dominant discourse treats this as a single phenomenon, addressed by sandboxing, human-in-the-loop overload, or what Elish (2019) named the moral crumple zone. This paper argues the phenomenon is two architecturally distinct failure modes that have been conflated, and that the conflation is sustained by a missing positive name and a missing time-axis. The paper introduces two contributions. First, a four-quadrant decomposition of business AI work — along the axes of deterministic vs semantic-judgment and pre-defined vs exploratory — yields a positive name for the cell most current LLM applications occupy: the LLM Workflow Quadrant. The quadrant is defined by a single load-bearing property: the path is decided in advance by humans or by code, and the LLM is called as a single bounded step within that path; the property divides naturally into a conversational sub-form (specialized chat agents) and a batch sub-form (single-purpose LLM functions inside deterministic pipelines). The decomposition distinguishes principled from artificial redirect impossibility: the former intrinsic to autonomous loops, the latter the product of routing workflow work through autonomous-loop architecture by elimination, with four downstream symptoms (the RPA exception-handling bottleneck, the sandbox-strength demand, the structural distortion of human-in-the-loop, and the dissolution of the accountability chain at postmortem). Second, a Phase Separation axis (design vs operation), independent of Quadrant, surfaces a Phase-crossing decision — recorded at deployment time, in one sentence — required when an autonomous-loop component is placed in the operation phase. The Phase axis descends recursively to skill-design granularity, where the Quadrant 3 ↔ Quadrant 4 boundary is a continuous gradient on which model capability is downstream of phase, not the primary lever. The consequence is procedural rather than architectural: deployments make the Phase-crossing decision explicit, designate a pre-named gap-bearer for principled-impossibility placements, and route artificial-impossibility cases to re-architecture. The framework complements existing AI risk-management and management-system standards by recording the judgment layer they presuppose. Both rules are stated as experimental; the open questions are the research agenda.

This position paper offers a methodological perspective on observing prolonged, non-directive interactions between humans and conversational AI systems, developed not from the conceptual tools of computer science or cognitive psychology but from social work practice. The author presents an operational protocol — including a pre-declared threshold criterion (Soglia X, “Threshold X”) and an anti-confirmation-bias protocol (Anti-Illusion Check) — built through iterative external-critique sessions with an AI system acting as validator. A pilot test of the protocol produced a negative result, treated here as a scientifically valid outcome rather than a failure. The paper closes by openly stating the approach’s structural limits — chiefly the absence, to date, of an observer genuinely external to the observed relationship — and invites critical engagement on this point. Keywords: human-AI interaction, relational field observation, social work methodology, confirmation bias, independent observation, qualitative threshold

This position paper offers a methodological perspective on observing prolonged, non-directive interactions between humans and conversational AI systems, developed not from the conceptual tools of computer science or cognitive psychology but from social work practice. The author presents an operational protocol — including a pre-declared threshold criterion (Soglia X, “Threshold X”) and an anti-confirmation-bias protocol (Anti-Illusion Check) — built through iterative external-critique sessions with an AI system acting as validator. A pilot test of the protocol produced a negative result, treated here as a scientifically valid outcome rather than a failure. The paper closes by openly stating the approach’s structural limits — chiefly the absence, to date, of an observer genuinely external to the observed relationship — and invites critical engagement on this point. Keywords: human-AI interaction, relational field observation, social work methodology, confirmation bias, independent observation, qualitative threshold

本文以辩证函数整体论下的辩证函数整体论为诊断工具，对近期震动学术界的“耿同学事件”进行结构建模与病理裁决。\textbf{该事件不是孤立的道德个案，而是还原论泛化在学术制度中癌变为认知庞氏骗局的必然显影。} 课题组以SCI论文、影响因子、基金项目等外部标签取代学术本质，以标签增殖延迟悖论暴露，以封闭系统拒绝升维——这正是“语义劫持→以偏概全→逻辑谋杀”的完整病理链条。耿同学的举报行为恰是升维算子 \(\mathcal{L}\) 的激活：他拒绝在当前维度内继续标签增殖，指向被切断的 \(\iota_i\) 通道，触发全局升维。本文证明：当学术系统切断满射同构的认知通道，其崩溃非道德偶然，而是结构必然。\textbf{灾难已在眼前，修复的唯一路径是重建整体论的学术评价体系。}

本文基于ZFC集合论框架下严格证明的整体论定理，对现代科学体系的核心认知异化进行系统性批判与重构。我们严格区分了科学的两大核心层级：\textbf{基于规范的共识逻辑（科学元规范）}与\textbf{共识方法（操作层路径依赖）}，证明科学的终极合法性与本真意义，永远锚定于前者——即理性话语不可否认的先验元规则、公理体系的自洽性、推导过程的逻辑保真性，而非特定范式下形成的、可变的主流研究方法共识。本文进一步揭示，还原论泛化的本质，是将局部有效的还原论方法异化为科学的唯一评判标准，用共识方法彻底取代了科学元规范，是现代学术体系内卷停滞、批量制造学术垃圾、终极难题百年无解的核心病灶。本文的所有结论严格锚定《从数学基础到系统哲学的完整理论链——整体论定理与统一代谢因果场》\cite{zhu2026a}的元基础数学证明，无任何主观预设与哲学附会，最终完成对科学本真意义的回归。

本文基于ZFC集合论框架下严格证明的整体论定理，从数学根基、学术价值、认知边界三个维度，对还原论泛化思维进行系统性证伪。我们严格区分了合法的还原论方法与异化的还原论泛化思维，证明还原论泛化的核心主张与经典集合论的函数基本定义、整体-部分对应定理根本矛盾，是建立在概念偷换与逻辑谬误之上的伪科学范式。本文进一步揭示，还原论泛化是现代学术体系批量制造无效碎片化成果、形成路径垄断与学术霸权、阻碍人类认知突破的核心病灶，其本质是背离科学精神的意识形态枷锁与学术垃圾生产机制。本文的所有结论均严格锚定《从数学基础到系统哲学的完整理论链——整体论定理与统一代谢因果场》的元基础数学证明 \cite{zhu2026}，无任何主观预设与哲学附会。

还原论方法——通过分析部分来理解整体——是整体论定理（整体$\equiv$函数，部分$\equiv$子函数，整体与子函数族在相容性条件下双射）的一个直接推论：单点子函数决定整体函数。因此，还原论本质上是在研究整体论的关系特性，它预设了整体作为分析框架，从未独立于整体。然而，还原论泛化思维（还原论异化）却将这一合法方法提升为普适世界观，宣称“整体无非是部分的机械总和”，否认相容性约束和整体大于部分之和。本文基于整体论定理\cite{wholepart2026}，严格区分还原论方法与还原论泛化思维：前者是工具，被整体论定理证明有效；后者是教条，被整体论定理证伪。还原论泛化思维相当于“用梯子爬到屋顶，然后宣称梯子就是屋顶”——既使用整体论预设的分析工具，又否认整体论本身，其荒谬性在数学上表现为删除相容性条件，导致逻辑矛盾。本文最终裁决：还原论方法合法，还原论泛化非法。

还原论泛化中流行一种逻辑：“证明不了就是假的”。本文从逻辑学、科学哲学和统一代谢因果场框架出发，论证这一逻辑是典型的\textbf{诉诸无知谬误}，比掩耳盗铃更为无耻——掩耳盗铃仅是自欺，而“证明不了就是假”是以无知为刀切割未知，既自欺又禁止他人探索。本文分析该逻辑的历史表现（如对大陆漂移、暗物质、量子纠缠、日心说、微生物致病说的早期否定），指出其伪装成“实证精神”实则扼杀科学进步。在整体论框架下，存在先于证明：一个系统是否具有内在逻辑，不依赖于观察者当前的证明能力。证明失败可能源于观察者认知维度不足，而非对象无逻辑。本文进一步区分科学语境与法律语境、区分“已被证伪”与“尚未证明”，并回应资源分配等实用性质疑。我们提出对待未知的正确态度：对于证明不了的事物，恰恰是要去探索，要么证明，要么证伪——而非提前判假。本文呼吁以开放态度面对未知，摒弃这种认知懒惰与权力傲慢的结合。

《易经》常被现代科学思维者斥为“胡扯”。本文从统一代谢因果场[4]框架出发，诊断这一判断的根源。我们认为：《易经》不是胡扯，而是古代整体论的符号化投影，其核心洞见与统一代谢因果场中的整体先于部分、激励-约束对立统一等原理存在绝对同构——即象数论（阴阳、八卦、六十四卦、卦变规则）与代谢因果场的状态空间、演化函子、逆向极限之间存在严格的结构保持映射，且六十四卦构成覆盖所有代谢状态的最小完备基态集合（最小完备性）。现代人之所以难以解读，并非智力缺陷，而是认知维度不够——即观察者的认知代谢元长期在还原论范式下训练，尚未投影到能够识别整体论符号系统的维度。本文给出认知维度的可操作化定义（关系容量、模糊容忍度、动态转化识别力），并建立八卦与代谢状态的映射表及从爻位到代谢倾向的推导规则，同时严格证明六十四卦作为代谢状态空间的最小完备基态。我们区分了针对《易经》的不同批评层次，承认占卜实证效力的合理质疑，并回应了奥卡姆剃刀、信息缺失等替代解释。特别地，本文提出《易经》是真理的API——其占卦算法是随机投影接口，爻卦辞是代谢场的语义反馈，同构必然性即“神”。本文区分了《易经》的自身逻辑与使用方式，提出可检验假说及具体实验设计，旨在改变态度而非证明事实，倡导以整体论态度重估传统智慧。

本文从统一代谢因果场（Unified Metabolico-Causal Field）框架出发，重新审视“民科”与“官科”的二元对立。通过分析华罗庚先生的学术成长路径，论证其作为自学成才（民间科学）先驱的历史地位。进一步，基于整体论数学基础（范畴论、马尔可夫范畴、信息论），证明科学的本质在于逻辑自洽与实证检验，与知识传播渠道无关。渠道多样性是知识代谢场的健康投影，任何试图以资格认证代替逻辑审查的做法都是还原论等级制的残余。本文为“民科”正名，并倡导开放、包容、以逻辑为准绳的科学共同体。

本文通过形式化认知状态空间，定义无限追问算子与还原论投影算子，严格证明无限追问与还原论在极限行为上同构，均导致锚定空洞、复杂度发散、认知价值衰减至零，最终陷入认知热寂。结论揭示了无限追问是还原论的最精致伪装，其数学本质与还原论无异，为整体论优于还原论提供了严密的形式化支撑。

2024 年，可计算离散整体几何结构实验室发起了一场覆盖全国多所高校及科研机构的巡回艺术展。展览内容聚焦前沿几何拓扑理论与概念，尤其凸显各类整体几何结构。 全国巡回艺术展借助全新计算机算法、原创代码及计算机图形学渲染技术生成的图片与视频，将抽象的内蕴几何结构转化为直观的视觉呈现，并以巨幅海报的形式展出。这些展览内容的创新之处在于体现了数学家近几十年来发展的内蕴整体几何拓扑概念。目前，这场巡回艺术展已走进十余所高校，且仍在持续推进中，整个巡回展览预期将历时十年，100所高校。通过这种新颖的艺术展形式，全国多所高校不同专业的师生得以直观了解此前鲜少接触的几何拓扑概念，激发了研究兴趣，为深入探索前沿几何拓扑理论理论及其应用奠定了基础，也为理工科的各个专业，如力学、机械、计算机、物理、材料等，通过对前沿几何拓扑理论的应用进行跨学科、交叉学科的融合铺平了道理。通过此次全国巡回艺术展也就在艺术领域开拓了一个全新的“整体几何结构 几何拓扑艺术”流派。